From 1ff5b3ed281755caf84ec7930019f184c30655a8 Mon Sep 17 00:00:00 2001 From: Yuta Saito Date: Mon, 19 Jan 2026 06:46:22 +0900 Subject: [PATCH] test: Fail MCP streaming test when LiteLLM logs errors during follow-up calls --- .../mcp_tests/test_aresponses_api_with_mcp.py | 165 ++++++++++-------- 1 file changed, 96 insertions(+), 69 deletions(-) diff --git a/tests/mcp_tests/test_aresponses_api_with_mcp.py b/tests/mcp_tests/test_aresponses_api_with_mcp.py index 2ec90ab5858..5871cece869 100644 --- a/tests/mcp_tests/test_aresponses_api_with_mcp.py +++ b/tests/mcp_tests/test_aresponses_api_with_mcp.py @@ -1,3 +1,4 @@ +import logging import os import sys import pytest @@ -667,7 +668,9 @@ async def test_streaming_mcp_events_validation(): pytest.param("anthropic/claude-4-5-haiku", id="anthropic"), ], ) -async def test_streaming_responses_api_with_mcp_tools(model: str): +async def test_streaming_responses_api_with_mcp_tools( + model: str, caplog: pytest.LogCaptureFixture +): """ Test the streaming responses API with MCP tools when using server_url="litellm_proxy" @@ -700,75 +703,99 @@ async def test_streaming_responses_api_with_mcp_tools(model: str): ] # Only mock the MCP-specific operations, let LLM responses be real - with patch.object(LiteLLM_Proxy_MCP_Handler, '_get_mcp_tools_from_manager', new_callable=AsyncMock) as mock_get_tools, \ - patch.object(LiteLLM_Proxy_MCP_Handler, '_execute_tool_calls', new_callable=AsyncMock) as mock_execute_tools: - - # Setup MCP mocks only - mock_get_tools.return_value = (mock_mcp_tools, ["litellm_proxy"]) - - # Create a dynamic mock that will match the actual tool call ID from the LLM response - def mock_execute_tool_calls_side_effect(tool_calls, user_api_key_auth): - """Mock function that returns results matching the actual tool call IDs from the LLM""" - results = [] - for tool_call in tool_calls: - # Extract call_id from the tool call - call_id = None - if isinstance(tool_call, dict): - call_id = tool_call.get("call_id") or tool_call.get("id") - elif hasattr(tool_call, 'call_id'): - call_id = tool_call.call_id - elif hasattr(tool_call, 'id'): - call_id = tool_call.id - - if call_id: - results.append({ - "tool_call_id": call_id, - "result": "LiteLLM is a unified interface for 100+ LLMs that translates inputs to provider-specific completion endpoints and provides consistent OpenAI-format output." - }) - return results - - mock_execute_tools.side_effect = mock_execute_tool_calls_side_effect - - # Make the actual call - LLM responses will be real - mcp_tool_config = cast(Any, { - "type": "mcp", - "server_url": "litellm_proxy", - "require_approval": "never" - }) - response = await litellm.aresponses( - model=model, - tools=[mcp_tool_config], - tool_choice="required", - input=[ + with caplog.at_level(logging.ERROR): + with patch.object( + LiteLLM_Proxy_MCP_Handler, + '_get_mcp_tools_from_manager', + new_callable=AsyncMock, + ) as mock_get_tools, patch.object( + LiteLLM_Proxy_MCP_Handler, + '_execute_tool_calls', + new_callable=AsyncMock, + ) as mock_execute_tools: + # Setup MCP mocks only + mock_get_tools.return_value = (mock_mcp_tools, ["litellm_proxy"]) + + # Create a dynamic mock that will match the actual tool call ID from the LLM response + def mock_execute_tool_calls_side_effect(tool_calls, user_api_key_auth): + """Mock function that returns results matching the actual tool call IDs from the LLM""" + results = [] + for tool_call in tool_calls: + # Extract call_id from the tool call + call_id = None + if isinstance(tool_call, dict): + call_id = tool_call.get("call_id") or tool_call.get("id") + elif hasattr(tool_call, 'call_id'): + call_id = tool_call.call_id + elif hasattr(tool_call, 'id'): + call_id = tool_call.id + + if call_id: + results.append( + { + "tool_call_id": call_id, + "result": "LiteLLM is a unified interface for 100+ LLMs that translates inputs to provider-specific completion endpoints and provides consistent OpenAI-format output.", + } + ) + return results + + mock_execute_tools.side_effect = mock_execute_tool_calls_side_effect + + # Make the actual call - LLM responses will be real + mcp_tool_config = cast( + Any, { - "role": "user", - "type": "message", - "content": "give me a TLDR of what BerriAI/litellm is about" - } - ], - stream=True - ) - - print(f"📋 Response type: {type(response)}") - assert hasattr(response, '__aiter__'), "Response should be an async streaming response" - - # Collect streaming chunks - chunks = [] - async for chunk in response: - chunks.append(chunk) - print(f"📦 Chunk type: {getattr(chunk, 'type', 'unknown')}") - - print(f"📊 Total chunks received: {len(chunks)}") - - # Verify MCP mocks were called (may be called multiple times in streaming) - assert mock_get_tools.call_count >= 1, f"Expected MCP tools to be fetched at least once, got {mock_get_tools.call_count}" - print(f"MCP tools fetched: {len(mock_mcp_tools)}") - - # Verify we got a response - assert response is not None - assert len(chunks) > 0, "Should have received streaming chunks" - - print("Basic streaming responses API with MCP tools test passed!") + "type": "mcp", + "server_url": "litellm_proxy", + "require_approval": "never", + }, + ) + response = await litellm.aresponses( + model=model, + tools=[mcp_tool_config], + tool_choice="required", + input=[ + { + "role": "user", + "type": "message", + "content": "give me a TLDR of what BerriAI/litellm is about", + } + ], + stream=True, + ) + + print(f"📋 Response type: {type(response)}") + assert hasattr(response, '__aiter__'), "Response should be an async streaming response" + + # Collect streaming chunks + chunks = [] + async for chunk in response: + chunks.append(chunk) + print(f"📦 Chunk type: {getattr(chunk, 'type', 'unknown')}") + + print(f"📊 Total chunks received: {len(chunks)}") + + # Verify MCP mocks were called (may be called multiple times in streaming) + assert ( + mock_get_tools.call_count >= 1 + ), f"Expected MCP tools to be fetched at least once, got {mock_get_tools.call_count}" + print(f"MCP tools fetched: {len(mock_mcp_tools)}") + + # Verify we got a response + assert response is not None + assert len(chunks) > 0, "Should have received streaming chunks" + + print("Basic streaming responses API with MCP tools test passed!") + + lite_errors = [ + record + for record in caplog.records + if record.levelno >= logging.ERROR + and ("LiteLLM" in record.name or "LiteLLM" in record.getMessage()) + ] + assert not lite_errors, "Unexpected LiteLLM errors: " + ", ".join( + record.getMessage() for record in lite_errors + ) @pytest.mark.asyncio